Cactus Compute publishes 12 papers in 2026, 5 at NeurIPS, focused on on-device efficiency
joemeno · x · 2026-11-20
Inference startup Cactus Compute published all 12 of its 2026 papers: 5 NeurIPS, 4 ICLR, 2 ICML, 1 arXiv.
The research focuses on small models and on-device inference efficiency, including:
- Temporal Mixture-of-Experts: serving memory that scales with active parameters
- Whole-model quantization below 4 bits per weight for on-device multimodal models
- Measuring attention's effective width in tokens
- Calibration-aware activation sparsity for instruction-tuned LLMs
- A blazing fast LM-head replacement
- GRPO-guided controllers for hyperparameter sweeps
- Adoption-gated continual learning for on-device tool-calling SLMs
Strategy: early results go to workshops, finished work is reserved for production blogs. Full list at cactuscompute.com/research.
More from Infra
- Liquid d1-2B decision model answers 19-30x more questions/min than same-size LLMs — mervenoyann · 2026-11-20
- Modular rebuilds MAX's Module API with PyTorch-like syntax and eager mode — clattner_llvm · 2026-11-20
- e-means lands across tusKANNy stack, cutting vector index clustering training time by 1.5-5x — CShorten30 · 2026-11-20
- AWS Lambda doubles function timeout to 30 minutes for on-demand Event Functions — DavidWells · 2026-11-20
- Microsoft taps Intel EMIB packaging for Azure silicon, locks in Diamond Rapids supply — Beth_Kindig · 2026-11-20
- Polymarket puts DOJ criminally charging Nvidia this year at just 5% — Polymarket · 2026-11-20